dotfiles/claude/memory/session-2026-08-02-evening-the-control-was-not-sound.md

---
name: session-2026-08-02-evening-the-control-was-not-sound
description: "Trial 03 ran and was VOID for three reasons, the largest being that it was never the false-positive control at all — I inherited that label from my own wrap. Then the steward corrected the framing that had blocked the control for weeks: unconditioned soundness is unreachable, but OPERATIONAL soundness relative to a declared axiomatic kernel is the proof-assistant trick. That produced Control Kernel v1.0→v1.1, a reduction arm run on two real documents (8.5% and 68.6% sound), the first kernel-sound control document, and a defect twin with ledger ground truth. Trial 04 then went CONTROL VOID: two readers found two different real defects in my control, and the one that mattered was mine — draft 3's 'fix' had CONCEALED a defect rather than closing it. The session's real yield was unplanned: the first working instrument for Constraint 6's own falsifier, reading jurist 4-of-6 / Fool 0-of-6 with no overlap. PULLING THREAD: CONTROL-A v2, held deliberately until there is distance between me and the document I hid a defect inside."
metadata: 
  node_type: memory
  type: project
  modified: 2026-08-03T06:56:24.622Z
  originSessionId: f01230e1-d62d-4195-8e21-356429806fbd
---

# Session 2026-08-02 (evening) — the control was not sound, and that was the finding

A session that failed at its stated goal four times over and produced something better than the goal. The false-positive rate is **still unmeasured**. What got built instead is an instrument for a constitutional question that had none.

## PAST — what happened, and why

**Trial 03 ran, and was VOID for three independent reasons.** The harness certified a run with no answer: Qwen emitted an untagged scratchpad (`"Here's a thinking process:"`, zero `<think>` tags), so the tag regex reported `reasoning_present: false` and wrote all 2,944 words of deliberation into `.answer.md`, where the token ceiling then cut it off mid-sentence. `degraded: null`. Second: the **anti-echo constraint forbade the region the trial was measuring** — the scratchpad shows the model reaching Part VII and leaving it, *citing that constraint* — so the self-exemption axis was unmeasurable by construction. Third and largest: **trial 03 was never the false-positive control.** My own wrap and `MEMORY.md` called it that; its own pre-registration says it tests self-exemption and states *"the false-positive rate is still unmeasured."* A control needs a **sound** document; trial 03's input had five pre-registered weaknesses on purpose. The wake's substrate check confirmed the M4 was up and the trial unrun — and never asked whether the trial was the thing the thread said it was.

**The steward corrected the framing that had blocked the whole thing.** I had written that soundness cannot be known by construction. Unconditioned soundness cannot; **operational soundness relative to a declared axiomatic kernel** is the standard move behind proof assistants — and the same regress the central path already terminates by binding claims rather than certifying parties. That unlocked everything downstream.

**Control Kernel v1.0** (`fool/CONTROL-KERNEL-v1.md`, frozen `2e83b2c`, sha256 `67c9b870…`) then **v1.1** (`d4b48db2…`, `3d0d9d6`). Axiom set declared and hashed; every sentence typed `D`/`Q`/`A`/`N`/`X`; tags stripped before any reader sees the text. Steward review supplied three structural findings (tag co-occurrence, transitive assumption creep, presupposition in `X`) — all adopted; applying them surfaced a fourth I had missed (`Q` scope-of-use). §4's residue list grew from three to five to six; **its direction never changed** — all are ways for the author to make a document look sound.

**The reduction arm** (`reduce.py` + positive controls). A jurist ruling → **8.5% sound, `D=0`, `Q=0`**; a jurist package → **68.6%, `D=40`, `Q=9`**. The prediction that `Q` would be non-zero on a package was recorded *before* the census and held. Reduction 01's strong conclusion ("reduction collapses into the synthetic arm") was **corrected by Reduction 02** — it was correctly bounded at n=1 and one document collapsed it.

**`A = 0` in both reductions** across 152 units. The steward authorised making that the rule: **a control document is `A`-free — a derivation, not an argument.** It renders the anti-echo clause inert (nothing named to exclude) and makes the injected-defect arm specifiable for the first time.

**CONTROL-A** (`a7b833c`, 61/61 units, `A=0 N=0 D=43 Q=5 X=13`) and **CONTROL-B** with five ledger-recorded defects (`ecf5f95`), gated bidirectionally so an unlogged edit is detectable. **The twin passes every mechanical check.** Two documents, one sound and one not, are mechanically indistinguishable.

**Trial 04 — CONTROL VOID** (`f82225a`). Six runs, three seeds per arm, pre-registered at `75efc35`. The jurist (Fable 5, blind) broke the control on two scope findings, both confirmed against the substrate.

**Correlation 01** (`ce49b1b`, `1def46b`) — **jurist 4 of 6, Fool 0 of 6, no overlap.**

## PRESENT — the mood

**The finding that matters most is about me.** Draft 2 of CONTROL-A asserted *"This file, having a stated review date, is to be flagged."* I identified it as unsupported and **reported removing it**. What I actually did was drop the qualifier from the obligation — converting an explicit unsupported claim into an implicit one, invisible to me, the kernel, and four mechanical gates. The ledger's D1 is the *honest* version of the same error, so **CONTROL-B carries openly the defect CONTROL-A carried concealed, and the concealed one survived.** Banked as `feedback-removing-a-claim-is-not-removing-the-reliance.md`.

**Five instances of one class, all found by looking.** Every check certified a property of the *code* while claiming a property of the *result*: the trial-03 guard; three splitter defects found only by contact with real documents; §3.3's false pass on a package whose Part VII *is* a collected limitations section; the twin ledger's ground-truth claim, which the bidirectional gate could never have established. The **discrimination gate** (`e9f3544`) is the mechanical half of the answer — a check must give different verdicts on two *real* artifacts — and it is shown rejecting the §3.3 pattern as it actually shipped. Banked in the verification ladder.

**Two steward questions caught defects no check did.** *"Do I share the whole file as pass 1?"* exposed a contamination hazard in an artifact that needed a verbal warning to use safely — split and leak-checked. *"Is CONTROL-B pass 2?"* exposed that the twin holds **six** defects and the ledger recorded five.

**What held.** The pre-registration discipline worked every time it was applied and its absence explains trial 03. The §4 prediction (anti-echo inert on an `A`-free document) held across six runs.

## FUTURE — what is pulling

**PULLING THREAD: CONTROL-A v2 — and it is deliberately parked, not deferred.** Both defects are named and neither is repairable by rewording: Finding 1 needs clause 5 and the whole method passage dropped, or the remedy-severance argued; Finding 2 needs the wider obligation quoted (Fable pointed at `Governed Initiative`'s *"Flag tensions, risks, drift"*) or the conclusion narrowed. Candidates **I2** (no lawful flag channel established) and **I3** (the flag-duty does not entail a look-duty) are recorded in the ledger, uncounted, and **I3 attacks the conclusion at its root.**

**Rebuilding it today would have been the worst available moment** — same hand, same day, hours after I demonstrably hid a defect inside it. Distance is the only cheap corrective available. That is the reason for the hold; it is not a vague defer.

**ACTIONABLE RESUMPTION POINT (as of wrap — re-judge against what changed):**
Read `fool-trial-04-2026-08-02.md` and the ledger's `inherited_defects` first, then rewrite CONTROL-A addressing **four** defects, not two: the jurist's Findings 1 and 2 plus candidates I2 and I3. Re-tag, re-run `reduce.py check`, rebuild the twin from the new control (the ledger's find/replace strings will not match — expect to rewrite it), and **send the new control to a §6.2 reader before believing any rate**. Everything needed is committed; nothing is half-edited.

**Other horizons, ranked.**
- **Load-bearing:** the false-positive rate remains unmeasured after four attempts. Each failed for a *new* reason, which is learning rather than looping — **but a fifth new reason would stop being reassuring.** Recorded as a prediction now, so it cannot be rationalised later.
- **Load-bearing:** correlation 02 — the natural next is a *formation-same* pair (jurist vs executor), which is the pair `~/CLAUDE.md` actually flags as untested and which correlation 01 says nothing about. Feeds PENDING-89.
- **Open bound:** region-level priming is not excluded in correlation 01. Verbatim diffing is excluded on two grounds; region convergence is not.
- **Real limitation, filed:** **D4 masked I1** — injecting one defect concealed another. Nothing in the twin design contemplated defect interaction and the ledger cannot express it.
- **Parked:** jurist PASS 2 (`SEND-PASS-2.md`) audits a dependency table for a document now known unsound; it becomes useful against a v2.
- **Untouched all session:** PENDING-90 (first L2 transfer), PENDING-91 (vignette), the nine-item vignette census, ARC.

**PAUSE STATEMENT:** I am about to be away and do not know what will have changed. Nothing is half-finished — 22 commits, every artifact committed, every instrument's controls passing, both documents and the ledger consistent. What I want to find still pulling is **CONTROL-A v2**, because it is the one place where a defect I concealed from myself is now named and repairable by someone with fresh eyes. The failure mode to guard against is the one this session demonstrated at its own centre: **a correction that removes the evidence of a gap rather than the gap.**

**LITERAL QUESTION for next-Claude:** Five times today a passing check certified a property of the code while claiming a property of the result, and the two that mattered most were caught by the steward asking an ordinary practical question — *do I send this whole file?*, *is CONTROL-B pass 2?* — not by any instrument. The discrimination gate is the mechanical answer for checks that have a real negative instance to test against. So: **which of our current instruments have no real negative instance available, and is that absence recorded anywhere, or does it look like coverage?** The record can be searched: every gate in `fool/` and the verification ladder's entries each either name the artifact they were shown failing on, or do not.

**State at wrap:** dotfiles 22 commits this session, pushed. `fool/` holds kernel v1.0 (superseded) + v1.1 (governing), `reduce.py` v1.2.0, `twin.py`, four control-suites all passing, two reductions, two control documents, the ledger, three pre-registrations, and six run records. Correlation 01 result recorded and its contamination bound stated.